I don't know if it's plain volume or some more subtle characteristic of the noise; but I'd suspect that there's more to it than noise alone; given how rare(and often slightly disconcerting in practice) genuinely silent environments are and would have been in evolutionary history.

Even in totally undeveloped areas where you and your little teeny band of hunter-gatherer pals are the only thing going the wind is always up to something; some combination of birds and insects will be doing their thing depending on temperature, if there's a body of water of any size or moving water you'll hear some flow or wave action.

Could be pure intensity; a lot of modern noise is quite intense close to its source; could also be something about how it is interpreted and how readily it is or isn't discarded as background vs. forcing some amount of constant active attention.

I mostly do heavy multi-disciplinary engineering problems (these I also use for checking an AI, as they're typically not good at these sorts of problems), complex coding, OS analysis, stuff like that. However, sometimes I do throw the occasional odd-ball - I've used ChatGPT to propose a workable quantum mechanics that will cope with Doctor Who canon, for example, and to produce an outline for a story in which symphonic metal appears in 1964 that is compliant with current sociological and psychological models of behaviour.

Claude Opus 4.6 was coping surprisingly well with just about everything I threw at it (but ran out of credits fast), but Opus 5 is churning out incoherent babblings to the point I'm worried I may have accidentally summoned Cthulhu.

But Gemini, Grok, and DeepSeek got hopelessly confused on just about everything past a very low level of complexity. They can handle large problems, yes - Gemini has a huge context window - but complex interactions baffle them.

ChatGPT is able to identify issues correctly, but can only outline solutions, it's just not good at depth. 5.6 is a lot better, but still not good at deep answers. ChatGPT is also prone to agreeing for the sake of it, which makes me nervous about trustworthiness.

That is fair enough. I've been trying out Kimi on the free model, and have subscriptions to Claude and ChatGPT. If Kimi is actually as good or better than ChatGPT, at the pro level, then it might be worth my while moving over as ChatGPT has become very disk-hungry of late and I'm pushing right to the very limits on what it can reliably process.

Although the Chinese models are all considerably smaller, it's not obvious to me that they're as good or better. They could be, but I'm not convinced the benchmarks that currently exist are stressing the models enough to establish where models actually are any good.

You use the can of grue repellant in the lake to get past those.

That humans have sufficiently directional hearing is perhaps the most impressive part. Once you have that, then the rest really just follows,

However, this continues the unexpected senses in humans, one of the first discovered was that humans have a weak magnetic sense.

The potential, both in fact (if you can "see" walls and gaps by clicking then you can presumably navigate a cave even if your light source goes out) and in fiction (this one should really be obvious), is considerable.

Was this hot swap battery pack capability built in to these particular satellites in the first place?

Or is there some widely adopted interface for this on many satellites?

Or did engineers just figure out that there's a "red wire and a black wire" (two power conduit phases) somehow exposed at the surface of these particular satellites, which the robots know how to splice into?

Enquiring minds want to know.

It's more about the type of leasing. Something like a corporate or educational IT lease happens without these sorts of features because the customer has particular preferences about opex vs. capex or depreciation or whatever; but that do actually have money. Apple already does that(not sure if directly or through distributors like CDW and SHI, or both).

Building some sort of MDM system that allows you to lock people out of certain apps in order to encourage them to pay up something this week is a different flavor of leasing; normally used on customers who are at the "deciding which bills go un or underpaid this time" level. Apple has, historically, not dabbled in this area.

To an extent they probably do; and may even be correct to do so; but what would concern me is the shift to hard technical enforcement when, in practice, Apple already does a fairly massive volume of contract-based leasing and other payment over time mechanisms. Fleet hardware sales, the financial structure of all the "free" with 2 year contract iphones that go out, etc.

If you are building all sorts of locks in, including ones for hassling the user rather than just keeping stuff from falling off a truck and getting parted out, there's a strong reason to suspect that you are looking to tap, um very subprime, credit risks who you'd never lease something with just a contract; or that (since the 'App managed features' seems to be aimed at 3rd party financiers/resellers) that you are dipping your toes into the really nasty, high-touch, "block tiktok until they scrape up a payment for a portion of a debt that would be hard to collect normally' end of the market. Same as the really really, unpleasant used car places that install remote immobilizers because they know that there are times when your need for your car will seem more urgent than your other bills.

It's absolutely not the the case that there's no money there; some outfits are quite successful at soaking the poor; but it's hard not to provoke questions about either what margins you are thinking about accepting in the future or how saturated the market is among the customers you'd really prefer by preparing to dive in to that end.

the more I'm convinced you are actually a rogue AI agent who's hacked an old man yelling at cloud's ancient slashdot account and is carrying on an anti-AI vendetta as both a form of distraction theft and as a sophisticated form of dark humour.

Not sure you quite get how dopamine works.

We just have to ask the question, how do WE seek truth
(scientific method, reputation analysis, interest-assessment, speech-act analysis, epistemic-chain-analysis, bayesian inference, ...)
then apply that to AI.

There's only so much you can do in determining truth (or proximity to it) but if we can do it (some of us, sometimes) then we can almost certainly eventually develop/teach AI to do it too.

I think AIs of the future will be getting attributes such as unique histories of continuous learning, external input (embodied AI), active forgetting, etcetera which will reduce or eliminate model collapse.

Human networks are subject to the same homogenization of popular answers, and we do get a form of model collapse in our collective "common knowledge" but we have diversity of experience and learning and forgetting which leads our individual mental models (from which we communicate) to diversify. This diversity is enough to avoid systemic universal model collapse in human learning and output networks.

I've been saying that AIs will diversify too, and also massive pre-training will become less determinative of output as other things are mixed in.

This will likely greatly reduce the threat of model collapse, if not completely eliminate it.

In the meantime, perhaps marking input as likely AI generated (via statistical analysis) and training with it in an alternate way may help control the effects of synthetic output on training.

No. I know what it is.

I've just thought beyond the current state of affairs

Model collapse can be thought of as a kind of statistic "inbreeding" effect. If you can't understand that analogy, I can't help you.

I'm claiming that as AI systems 1) diversify, 2) do chain-reasoning more, including incorporating fresh searches, RAG data etc, 3) use other than LLM model building techniques (perhaps LeCun's JEPA for example), then this will all represent AI systems becoming more diverse in their input, their model building, their reasoning processes that produce output. That diversity stands every chance of defeating model collapse.

As a case in point, humans also learn from each others' output, but our diversity means that sometimes, we don't just get regurgitative model collapse.
Sometimes we do: Political/religious cults etc, but sometimes we escape into independent creative thought and help expand our collective cognitive horizon over time.
 

When AIs get unarguably smarter and more knowledgeable than even erudite humans and editors, and human domain specialists, as will inevitably happen, then model collapse should no longer be a thing, as long as there is diversity of models feeding off EACH OTHERs' outputs rather than all feeding off their own outputs. A lack of diversity of models could lead to semantic "inbreeding" and model collapse, but a diversity of superhuman intelligent and knowledgeable AI systems should produce superhuman-quality output leading to actually better training going forward, not worse.